Development of Universal and Scalable Gaussian Process Machine Learning Potential

Abstract: 

Machine learning has transformed numerous scientific disciplines by tackling key challenges. Machine learning potential is one of key applications offering a promising avenue for surmounting the stringent limitations on size and speed of ab initio simulations [1]. Our group focuses on developing transferable and scalable sparse Gaussian process regression (SGPR) based machine learning potentials [2,3,4,5]. In principle, the SGPR machine learning potential can be expanded into a general potential for all known compounds. The SGPR machine learning potential has been developed for a variety of materials, including Li solid electrolytes [2], hydrocarbons [3], Li-battery cathodes [4], and aqueous solutions [5]. Currently, machine learning potentials only predict ab initio energy based on geometric information, which lacks crucial electronic details such as charge states, excited states, and spin interactions. To address this limitation, we incorporate the missing electronic structure information, including electron transfer [5] and spin-spin interactions [6], enhancing its predictive capabilities of the SGPR potential. The ultimate goal is to build and apply this general-purpose ab initio machine learning potential to investigate various complex reactions such as catalysis, electron transfer, protein folding, and adsorption phenomena on surfaces. 

References 

[1] C. W. Myung, B. Hirshberg, M. Parrinello, Phys. Rev. Lett. 128, 045301 (2022); B. X. Shi, A. Michaelides, C. W. Myung, Establishing the gold-standard for oxide-supported nanoclusters: Coupled cluster benchmarks of coinage metal structures on MgO, to be submitted (2023). 

[2] A. Hajibabaei, C. W. Myung, K. S. Kim Phys. Rev. B 103, 214102 (2021).; C. W. Myung et al. Adv. Energy Mater. 12, 2202279 (2022). 

[3] M. Ha et al., ACS Phys. Chem Au, 2, 260?264 (2022); S. Y. Willow, G. S. Kim, M. Ha, A. Hajibabaei, C. W. Myung, A Scalable Bayesian Committee Machine Machine Learning Potential for Hydrocarbons. to be submitted (2023). 

[4] M. Ha et al. Adv. Energy Mater. 12, 2201497 (2022). 

[5] C. W. Myung et al, Active multi-task sparse Gaussian process regression machine learning potential: Redox chemistry of aqueous solutions, to be submitted, (2023). 

[6] T. H. Park, D. C. Ryu, C. J. Kang, C. W. Myung, Spin-dependent sparse Gaussian regression potential using renormalized spin-exchange function, in preparation, (2023).

20230921_대학원세미나_명창우 교수.pdf